Glow-TTS: A Generative Flow for Text-to-Speech via Monotonic Alignment Search

Understand

Recently, text-to-speech (TTS) models such as FastSpeech and ParaNet have been proposed to generate mel-spectrograms from text in parallel.

• Despite the advantage, the parallel TTS models cannot be trained without guidance from autoregressive TTS models as their external aligners.
• In this work, we propose Glow-TTS, a flow-based generative model for parallel TTS that does not require any external aligner.
• By combining the properties of flows and dynamic programming, the proposed model searches for the most probable monotonic alignment between text and the latent representation of speech on its own.